Predicting phenotypes from genetic, environment, management, and historical data using CNNs

Abstract

Predicting phenotypes from genetic (G), environmental (E), and management (M) conditions is a long-standing challenge with implications to agriculture, medicine, and conservation. Most methods reduce the factors in a dataset (feature engineering) in a subjective and potentially oversimplified manner. Convolutional Neural Networks (CNN) can overcome this by allowing the data itself to determine which factors are most important. CNN models were developed for predicting agronomic yield from a combination of replicated trials and historical yield survey data. The results were more accurate than standard methods when tested on heldout G, E, and M data (r=0.5 vs r=0.4), and performed slightly worse than standard methods when only G was held out (r=0.74 vs r=0.78). Pre-training on historical data increased accuracy by 1-36% compared to trial data alone. Saliency map analysis indicated the CNN has "learned" to prioritize many factors of known agricultural importance.

Yield Prediction Through Integration of Genetic, Environment, and Management Data Through Deep Learning

“…Proceeding in this manner rather than selecting observations at random for the testing set further reduces an already small number of weather and management conditions. Incorporating historical data (Washburn et al . 2021) or expanding the dataset to include data from other sources represent two possible avenues to incorporate a greater diversity of weather and management conditions without compromising the testing set.…”

“…Although we aimed to broaden the range of possible architectures relative to previous modeling on G2F data (Washburn et al . 2021), we constrained the overall structure to processing each tensor individually then allowing for interactions between the final layer of each module.…”

“…We use one dimensional convolutional layers to capture the time dependent features of weather data, which have previously been used in yield prediction for this task (Khaki et al . 2020; Washburn et al . 2021).…”

“…2001) this field is rapidly developing, with advances in theory, software, and hardware enabling deeper and more accurate networks. Several recent studies have applied these methods with a relatively large dataset either with (Washburn et al . 2021) or without (Khaki et al .…”

“…We begin by detailing a workflow that incorporates a wider number of years from the Genomes to Fields Initiative in than previous studies (Rogers et al . 2021; Washburn et al . 2021; Rogers and Holland 2021), while also limiting the effect of errant and absent measurements.…”